Operations on Spherical Tensors

using the se3cnn repository

tutorial by: Tess E. Smidt

code by:

@misc{mario_geiger_2019_3348277,
  author       = {Mario Geiger and
                  Tess Smidt and
                  Wouter Boomsma and
                  Maurice Weiler and
                  Michał Tyszkiewicz and
                  Jes Frellsen and
                  Benjamin K. Miller},
  title        = {mariogeiger/se3cnn: Point cloud support},
  month        = jul,
  year         = 2019,
  doi          = {10.5281/zenodo.3348277},
  url          = {https://doi.org/10.5281/zenodo.3348277}
}

We're going to use the SphericalTensor class like we did in data_types.ipynb.

import torch
from spherical import SphericalTensor

torch.set_default_dtype(torch.float64)

Spherical tensors can be added.

Rs = [(1, 1)]
sum_Ls = sum((2 * L + 1) for mult, L in Rs for _ in range(mult))
signal_1 = torch.zeros(sum_Ls)
signal_1[0] = 1.  # y
signal_2 = torch.zeros(sum_Ls)
signal_2[2] = 1.  # x
sphten_1 = SphericalTensor(signal_1, Rs)
sphten_2 = SphericalTensor(signal_2, Rs)

import plotly
from plotly.subplots import make_subplots

n = 50

def plot_operation(input1, input2, output):
    rows = 1
    cols = 3
    specs = [[{'is_3d': True} for i in range(cols)]
             for j in range(rows)]
    fig = make_subplots(rows=rows, cols=cols, specs=specs)
    for i, sphten in enumerate([input1, input2, output]):
        trace = sphten.plot(relu=False, n=n)
        trace.showscale = False
        fig.add_trace(trace, row=1, col=i + 1)
    fig.update_layout(scene_aspectmode='data')
    return fig

new_sphten = sphten_1 + sphten_2
# plots functions proportional to y, x, and (x + y)
fig = plot_operation(sphten_1, sphten_2, new_sphten)
fig.show()

We can compute the dot product of two spherical tensors

dot_product = sphten_1 * sphten_2 # These functions are orthogonal
print(dot_product)

dot_product = sphten_1 * sphten_1 # These functions are identical
print(dot_product)

tensor(0.)
tensor(1.)

We CANNOT multiply two spherical tensors, but we can compute their tensor product and use Clebsch-Gordon coeffients to combine two tensor indices into one

new_sphten = sphten_1 @ sphten_2
print("input1 Rs", sphten_1.Rs)
print("input2 Rs", sphten_2.Rs)
print("output Rs", new_sphten.Rs)
print("")
# plots functions proportional to y, x, and 1 + xy
print("Now we have contributions to z (cross product) and xy (outer product).")
fig = plot_operation(sphten_1, sphten_2, new_sphten)
print("SH:", "  1      y      z      x      xy     yz     *      zx     %",)
print("new", new_sphten.signal.numpy().round(3))
print("* == 2z^2 - x^2 - y^2")
print("% == x^2 - y^2")
fig.show()

input1 Rs [(1, 1)]
input2 Rs [(1, 1)]
output Rs [(1, 0), (1, 1), (1, 2)]

Now we have contributions to z (cross product) and xy (outer product).
SH:   1      y      z      x      xy     yz     *      zx     %
new [ 0.    -0.     0.707  0.     0.707  0.    -0.    -0.    -0.   ]
* == 2z^2 - x^2 - y^2
% == x^2 - y^2